kernel/memremap.c

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 /* Copyright(c) 2015 Intel Corporation. All rights reserved. */
   3 #include <linux/radix-tree.h>
   4 #include <linux/device.h>
   5 #include <linux/types.h>
   6 #include <linux/pfn_t.h>
   7 #include <linux/io.h>
   8 #include <linux/kasan.h>
   9 #include <linux/mm.h>
  10 #include <linux/memory_hotplug.h>
  11 #include <linux/swap.h>
  12 #include <linux/swapops.h>
  13 #include <linux/wait_bit.h>
  14
  15 static DEFINE_MUTEX(pgmap_lock);
  16 static RADIX_TREE(pgmap_radix, GFP_KERNEL);
  17 #define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
  18 #define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
  19
  20 static unsigned long order_at(struct resource *res, unsigned long pgoff)
  21 {
  22         unsigned long phys_pgoff = PHYS_PFN(res->start) + pgoff;
  23         unsigned long nr_pages, mask;
  24
  25         nr_pages = PHYS_PFN(resource_size(res));
  26         if (nr_pages == pgoff)
  27                 return ULONG_MAX;
  28
  29         /*
  30          * What is the largest aligned power-of-2 range available from
  31          * this resource pgoff to the end of the resource range,
  32          * considering the alignment of the current pgoff?
  33          */
  34         mask = phys_pgoff | rounddown_pow_of_two(nr_pages - pgoff);
  35         if (!mask)
  36                 return ULONG_MAX;
  37
  38         return find_first_bit(&mask, BITS_PER_LONG);
  39 }
  40
  41 #define foreach_order_pgoff(res, order, pgoff) \
  42         for (pgoff = 0, order = order_at((res), pgoff); order < ULONG_MAX; \
  43                         pgoff += 1UL << order, order = order_at((res), pgoff))
  44
  45 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE)
  46 vm_fault_t device_private_entry_fault(struct vm_area_struct *vma,
  47                        unsigned long addr,
  48                        swp_entry_t entry,
  49                        unsigned int flags,
  50                        pmd_t *pmdp)
  51 {
  52         struct page *page = device_private_entry_to_page(entry);
  53
  54         /*
  55          * The page_fault() callback must migrate page back to system memory
  56          * so that CPU can access it. This might fail for various reasons
  57          * (device issue, device was unsafely unplugged, ...). When such
  58          * error conditions happen, the callback must return VM_FAULT_SIGBUS.
  59          *
  60          * Note that because memory cgroup charges are accounted to the device
  61          * memory, this should never fail because of memory restrictions (but
  62          * allocation of regular system page might still fail because we are
  63          * out of memory).
  64          *
  65          * There is a more in-depth description of what that callback can and
  66          * cannot do, in include/linux/memremap.h
  67          */
  68         return page->pgmap->page_fault(vma, addr, page, flags, pmdp);
  69 }
  70 EXPORT_SYMBOL(device_private_entry_fault);
  71 #endif /* CONFIG_DEVICE_PRIVATE */
  72
  73 static void pgmap_radix_release(struct resource *res, unsigned long end_pgoff)
  74 {
  75         unsigned long pgoff, order;
  76
  77         mutex_lock(&pgmap_lock);
  78         foreach_order_pgoff(res, order, pgoff) {
  79                 if (pgoff >= end_pgoff)
  80                         break;
  81                 radix_tree_delete(&pgmap_radix, PHYS_PFN(res->start) + pgoff);
  82         }
  83         mutex_unlock(&pgmap_lock);
  84
  85         synchronize_rcu();
  86 }
  87
  88 static unsigned long pfn_first(struct dev_pagemap *pgmap)
  89 {
  90         const struct resource *res = &pgmap->res;
  91         struct vmem_altmap *altmap = &pgmap->altmap;
  92         unsigned long pfn;
  93
  94         pfn = res->start >> PAGE_SHIFT;
  95         if (pgmap->altmap_valid)
  96                 pfn += vmem_altmap_offset(altmap);
  97         return pfn;
  98 }
  99
 100 static unsigned long pfn_end(struct dev_pagemap *pgmap)
 101 {
 102         const struct resource *res = &pgmap->res;
 103
 104         return (res->start + resource_size(res)) >> PAGE_SHIFT;
 105 }
 106
 107 static unsigned long pfn_next(unsigned long pfn)
 108 {
 109         if (pfn % 1024 == 0)
 110                 cond_resched();
 111         return pfn + 1;
 112 }
 113
 114 #define for_each_device_pfn(pfn, map) \
 115         for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
 116
 117 static void devm_memremap_pages_release(void *data)
 118 {
 119         struct dev_pagemap *pgmap = data;
 120         struct device *dev = pgmap->dev;
 121         struct resource *res = &pgmap->res;
 122         resource_size_t align_start, align_size;
 123         struct page *first_page;
 124         unsigned long pfn;
 125         int nid;
 126
 127         pgmap->kill(pgmap->ref);
 128         for_each_device_pfn(pfn, pgmap)
 129                 put_page(pfn_to_page(pfn));
 130
 131         /* pages are dead and unused, undo the arch mapping */
 132         align_start = res->start & ~(SECTION_SIZE - 1);
 133         align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
 134                 - align_start;
 135
 136         /* make sure to access a memmap that was actually initialized */
 137         first_page = pfn_to_page(pfn_first(pgmap));
 138
 139         nid = page_to_nid(first_page);
 140
 141         mem_hotplug_begin();
 142         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 143                 pfn = align_start >> PAGE_SHIFT;
 144                 __remove_pages(pfn, align_size >> PAGE_SHIFT, NULL);
 145         } else {
 146                 arch_remove_memory(nid, align_start, align_size,
 147                                 pgmap->altmap_valid ? &pgmap->altmap : NULL);
 148                 kasan_remove_zero_shadow(__va(align_start), align_size);
 149         }
 150         mem_hotplug_done();
 151
 152         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 153         pgmap_radix_release(res, -1);
 154         dev_WARN_ONCE(dev, pgmap->altmap.alloc,
 155                       "%s: failed to free all reserved pages\n", __func__);
 156 }
 157
 158 /**
 159  * devm_memremap_pages - remap and provide memmap backing for the given resource
 160  * @dev: hosting device for @res
 161  * @pgmap: pointer to a struct dev_pagemap
 162  *
 163  * Notes:
 164  * 1/ At a minimum the res, ref and type members of @pgmap must be initialized
 165  *    by the caller before passing it to this function
 166  *
 167  * 2/ The altmap field may optionally be initialized, in which case altmap_valid
 168  *    must be set to true
 169  *
 170  * 3/ pgmap->ref must be 'live' on entry and will be killed at
 171  *    devm_memremap_pages_release() time, or if this routine fails.
 172  *
 173  * 4/ res is expected to be a host memory range that could feasibly be
 174  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 175  *    this is not enforced.
 176  */
 177 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
 178 {
 179         resource_size_t align_start, align_size, align_end;
 180         struct vmem_altmap *altmap = pgmap->altmap_valid ?
 181                         &pgmap->altmap : NULL;
 182         struct resource *res = &pgmap->res;
 183         unsigned long pfn, pgoff, order;
 184         pgprot_t pgprot = PAGE_KERNEL;
 185         int error, nid, is_ram;
 186         struct dev_pagemap *conflict_pgmap;
 187
 188         if (!pgmap->ref || !pgmap->kill)
 189                 return ERR_PTR(-EINVAL);
 190
 191         align_start = res->start & ~(SECTION_SIZE - 1);
 192         align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
 193                 - align_start;
 194         align_end = align_start + align_size - 1;
 195
 196         conflict_pgmap = get_dev_pagemap(PHYS_PFN(align_start), NULL);
 197         if (conflict_pgmap) {
 198                 dev_WARN(dev, "Conflicting mapping in same section\n");
 199                 put_dev_pagemap(conflict_pgmap);
 200                 return ERR_PTR(-ENOMEM);
 201         }
 202
 203         conflict_pgmap = get_dev_pagemap(PHYS_PFN(align_end), NULL);
 204         if (conflict_pgmap) {
 205                 dev_WARN(dev, "Conflicting mapping in same section\n");
 206                 put_dev_pagemap(conflict_pgmap);
 207                 return ERR_PTR(-ENOMEM);
 208         }
 209
 210         is_ram = region_intersects(align_start, align_size,
 211                 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 212
 213         if (is_ram != REGION_DISJOINT) {
 214                 WARN_ONCE(1, "%s attempted on %s region %pr\n", __func__,
 215                                 is_ram == REGION_MIXED ? "mixed" : "ram", res);
 216                 error = -ENXIO;
 217                 goto err_array;
 218         }
 219
 220         pgmap->dev = dev;
 221
 222         mutex_lock(&pgmap_lock);
 223         error = 0;
 224
 225         foreach_order_pgoff(res, order, pgoff) {
 226                 error = __radix_tree_insert(&pgmap_radix,
 227                                 PHYS_PFN(res->start) + pgoff, order, pgmap);
 228                 if (error) {
 229                         dev_err(dev, "%s: failed: %d\n", __func__, error);
 230                         break;
 231                 }
 232         }
 233         mutex_unlock(&pgmap_lock);
 234         if (error)
 235                 goto err_radix;
 236
 237         nid = dev_to_node(dev);
 238         if (nid < 0)
 239                 nid = numa_mem_id();
 240
 241         error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(align_start), 0,
 242                         align_size);
 243         if (error)
 244                 goto err_pfn_remap;
 245
 246         mem_hotplug_begin();
 247
 248         /*
 249          * For device private memory we call add_pages() as we only need to
 250          * allocate and initialize struct page for the device memory. More-
 251          * over the device memory is un-accessible thus we do not want to
 252          * create a linear mapping for the memory like arch_add_memory()
 253          * would do.
 254          *
 255          * For all other device memory types, which are accessible by
 256          * the CPU, we do want the linear mapping and thus use
 257          * arch_add_memory().
 258          */
 259         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 260                 error = add_pages(nid, align_start >> PAGE_SHIFT,
 261                                 align_size >> PAGE_SHIFT, NULL, false);
 262         } else {
 263                 error = kasan_add_zero_shadow(__va(align_start), align_size);
 264                 if (error) {
 265                         mem_hotplug_done();
 266                         goto err_kasan;
 267                 }
 268
 269                 error = arch_add_memory(nid, align_start, align_size, altmap,
 270                                 false);
 271         }
 272
 273         if (!error) {
 274                 struct zone *zone;
 275
 276                 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
 277                 move_pfn_range_to_zone(zone, align_start >> PAGE_SHIFT,
 278                                 align_size >> PAGE_SHIFT, altmap);
 279         }
 280
 281         mem_hotplug_done();
 282         if (error)
 283                 goto err_add_memory;
 284
 285         for_each_device_pfn(pfn, pgmap) {
 286                 struct page *page = pfn_to_page(pfn);
 287
 288                 /*
 289                  * ZONE_DEVICE pages union ->lru with a ->pgmap back
 290                  * pointer.  It is a bug if a ZONE_DEVICE page is ever
 291                  * freed or placed on a driver-private list.  Seed the
 292                  * storage with LIST_POISON* values.
 293                  */
 294                 list_del(&page->lru);
 295                 page->pgmap = pgmap;
 296                 percpu_ref_get(pgmap->ref);
 297         }
 298
 299         error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
 300                         pgmap);
 301         if (error)
 302                 return ERR_PTR(error);
 303
 304         return __va(res->start);
 305
 306  err_add_memory:
 307         kasan_remove_zero_shadow(__va(align_start), align_size);
 308  err_kasan:
 309         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 310  err_pfn_remap:
 311  err_radix:
 312         pgmap_radix_release(res, pgoff);
 313  err_array:
 314         pgmap->kill(pgmap->ref);
 315         return ERR_PTR(error);
 316 }
 317 EXPORT_SYMBOL_GPL(devm_memremap_pages);
 318
 319 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 320 {
 321         /* number of pfns from base where pfn_to_page() is valid */
 322         return altmap->reserve + altmap->free;
 323 }
 324
 325 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 326 {
 327         altmap->alloc -= nr_pfns;
 328 }
 329
 330 /**
 331  * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
 332  * @pfn: page frame number to lookup page_map
 333  * @pgmap: optional known pgmap that already has a reference
 334  *
 335  * If @pgmap is non-NULL and covers @pfn it will be returned as-is.  If @pgmap
 336  * is non-NULL but does not cover @pfn the reference to it will be released.
 337  */
 338 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 339                 struct dev_pagemap *pgmap)
 340 {
 341         resource_size_t phys = PFN_PHYS(pfn);
 342
 343         /*
 344          * In the cached case we're already holding a live reference.
 345          */
 346         if (pgmap) {
 347                 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
 348                         return pgmap;
 349                 put_dev_pagemap(pgmap);
 350         }
 351
 352         /* fall back to slow path lookup */
 353         rcu_read_lock();
 354         pgmap = radix_tree_lookup(&pgmap_radix, PHYS_PFN(phys));
 355         if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
 356                 pgmap = NULL;
 357         rcu_read_unlock();
 358
 359         return pgmap;
 360 }
 361 EXPORT_SYMBOL_GPL(get_dev_pagemap);
 362
 363 #ifdef CONFIG_DEV_PAGEMAP_OPS
 364 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
 365 EXPORT_SYMBOL(devmap_managed_key);
 366 static atomic_t devmap_enable;
 367
 368 /*
 369  * Toggle the static key for ->page_free() callbacks when dev_pagemap
 370  * pages go idle.
 371  */
 372 void dev_pagemap_get_ops(void)
 373 {
 374         if (atomic_inc_return(&devmap_enable) == 1)
 375                 static_branch_enable(&devmap_managed_key);
 376 }
 377 EXPORT_SYMBOL_GPL(dev_pagemap_get_ops);
 378
 379 void dev_pagemap_put_ops(void)
 380 {
 381         if (atomic_dec_and_test(&devmap_enable))
 382                 static_branch_disable(&devmap_managed_key);
 383 }
 384 EXPORT_SYMBOL_GPL(dev_pagemap_put_ops);
 385
 386 void __put_devmap_managed_page(struct page *page)
 387 {
 388         int count = page_ref_dec_return(page);
 389
 390         /*
 391          * If refcount is 1 then page is freed and refcount is stable as nobody
 392          * holds a reference on the page.
 393          */
 394         if (count == 1) {
 395                 /* Clear Active bit in case of parallel mark_page_accessed */
 396                 __ClearPageActive(page);
 397                 __ClearPageWaiters(page);
 398
 399                 mem_cgroup_uncharge(page);
 400
 401                 page->pgmap->page_free(page, page->pgmap->data);
 402         } else if (!count)
 403                 __put_page(page);
 404 }
 405 EXPORT_SYMBOL(__put_devmap_managed_page);
 406 #endif /* CONFIG_DEV_PAGEMAP_OPS */